ModelCard for UnGPT-v1

Model Details

• Name: UnGPT-v1
• Foundation Model: Mistral v0.3 (7B parameters)
• Recommended Context Length: 16k tokens
• Fine-tuning Methodology: LoRA-based training with Odds Ratio Preference Optimization method, using a combination of ebooks and synthetic data.

Usage Instructions

Use the Alpaca format for prompts:

### Instruction:
{instruction}

### Input:
{input}

### Response:

Example prompts

For instructions, it is not recommended to deviate from the provided examples. For the input, a minimum is 10 sentences, but more can work as the model can handle longer context sizes (Thanks to the Mistral 7B v0.3 base model).

1. Completion Prompt:

   ### Instruction:
   Continue writing the story while retaining writing style. Write about 10 sentences.
   
   ### Input:
   It was a dark and stormy night...
   
   ### Response:
   

2. Fill-in-the-middle Prompt:

   ### Instruction:
   Fill in the missing part of the story ({{FILL_ME}}) with about 10 sentences while retaining the writing style.
   
   ### Input:
   The bus was speeding down the road, cops chasing after it. 
   {{FILL_ME}}
   She woke up to find herself in an unfamiliar room...
   
    ### Response:
   

Dataset Preparation

For dataset acquisition and cleanup please refer steps 1 and 2 of my text-completion example, molbal/llm-text-completion-finetune.

Chunking: Split texts into chunks based on sentence boundaries, aiming for 100 sentences per example.

• For completion examples, 90 sentences were used as input, 10 sentences as response.
• For fill-in-the-middle examples, 80 + 10 sentences as input (before and after the {{FILL_ME}} placeholder, respectively), and 10 sentences as response.

The beauty of the ORPO method is that for a single prompt we can set both a positive and a negative example. I wanted the model to avoid 'GPTisms' so I had gpt4o-mini generate answers both for completion and FOM tasks and added them as a neative example.

The dataset used is ~15k examples, each approximately 9000 characters long including input, accepted and refused response. (Note these are characters not tokens)

Training setup

• Fine-tuned the Mistral v0.3 foundation model using Unsloth and ORPO trainer.

• Training configuration:

  • Batch size: 1
  • Gradient accumulation steps: 4
  • Learning rate scheduler type: Linear
  • Optimizer: AdamW (8-bit)
  • Number of training epochs: 1

• Hardware

  • I used GPU accelerated containers from the provider vast.ai (My referral link: https://cloud.vast.ai/?ref_id=123492 ) and executed training for ~8 hours on a single RTX 4090.

• Training costs

  • ~5€ for renting a GPU pod (+15€ in unsuccessful attempts)
  • ~5€ in OpenAI API costs for generating refusals

Licensing and Citation

• License: This model is licensed under the Apache License 2.0.
• Citation:

@misc{ungpt-v1,
  author = Bálint Molnár-Kaló,
  title = {UnGPT-v1: A Fine-tuned Mistral Model for Story Continuation},
  howpublished = {\url{https://huggingface.co/models/molbal/UnGPT-v1}},
  year = 2024
}